Amp now uses Claude Opus 5.5 as the default medium mode, replacing GPT-5.6 Sol and solving 65% of tasks at 10% lower cost. Opus 5.5 responds to mid-work corrections—add constraints, point out errors, or redirect without restarting. The model runs at high reasoning effort by design; beyond that, it overthinks and costs far more for worse results. Use ultra with Fable 5.1 for exploratory work with many unknowns, or high with GPT-6 Astra for decisions requiring taste and judgment.